Context

Most of the time, accurate diagnosis of cancer requires a histological examination. It consists of analyzing a tissue sample so as to confirm the presence of a tumor, qualify the type of lesion detected and, thus, adjust the therapy.

At primaa, we have developed deep learning based methods to classify lesions, detect biomarkers and segment regions of interests. Training models from scratch with no a priori knowledge requires a high quantity of data. A classic approach to overcome this limitation is to use models pre-trained on ImageNet classification tasks, when available. Another promising way is to use backbone models trained in an unsupervised manner. Several authors have applied such techniques for histology purposes (see [1], [2]).
Once a backbone is available, it can be used to support various applications such as feature encoding for further classification or segmentation tasks, few shot learning, weak supervised learning…

In this internship, we will explore two main paths:

• Train a model backbone in an unsupervised manner.
• Use a given, frozen backbone to perform classification and segmentation tasks by adding small, trainable heads.

[1] :Self-Supervised Vision Transformers Learn Visual Concepts in Histopathology, Chen, Richard J and Krishnan, Rahul G, Learning Meaningful Representations of Life, NeurIPS 2021
[2] :A Simple Framework for Contrastive Learning of Visual Representations, Chen, Kornblith, Norouzi, Hinton, ICML’2020

Objectives

• Select relevant approaches to train a model in an unsupervised manner for histology purposes, based on a literature review.

• Use a given backbone to perform classification and segmentation tasks:

  • Design data augmentation strategies in the feature space.
  • Design relevant classification/segmentation heads.
  • Explore Few-Shot Learning based techniques.

Practical information

• Remuneration : 1000€/month + 50% navigo refund
• Primaa is based in Paris, 2ème arrondissement
• possible occasional remote

• Applied mathematics student.
• Six months internship (start between march and May 2023.)
• Knowledge in python.
• Knowledge in computer vision (image processing, deep learning) : tensorflow and/or pytorch, scikit-image ….

Meeting with two data team members